The Self-Organizing Map (SOM) algorithm has attracted an ever increasing amount of interest among researches and practitioners in a wide variety of elds The SOM and a variant of it, the LVQ, have been analyzed extensively, a number of variants of them have been developed and, perhaps most notably, they have been applied extensively within elds ranging from engineering sciences to medicine, biology, and economics We have collected a comprehensive list of 3343 scienti c papers that use the algorithms, have bene ted from them, or contain analyses of them The list is intended to serve as a source for literature surveys We have provided both a thematic and a keyword index to help nding articles of interest

Bibliography of Self-Organizing Map (SOM) Papers: 1981-1997

We present a modification of Kohonens SOFMin order to simulate cortical plasticity induced by coactivation patterns. This is accomplished by introducing a probabilistic mode of stimulus presentation and by substituting the winner-takes- all mechanism by selecting the winner from a set of best matching neurons.

Modification of Kohonen's SOFM to Simulate Cortical Plasticity Induced by Coactivation Input Patterns

In this paper we used WPT (wavelet packet transform) and neural classifier SVM (support vector machine) to recognize spoken digits from 0 to 9 in Brazilian Portuguese. The main objective this work is to find out the Wavelet mother that better represents the speech signal in Brazilian Portuguese. The results obtained were compared with MFCC (Mel frequency cepstral coefficients). We carried out sixteen experiments with different Wavelets in dependent- case and four experiments in independent-case. The database was recorded in three months with 82 eighteen-to- forty years old male speakers. The SVM was used as a classifier in a "one vs. all" strategy. Best results have been obtained using Wavelets Daubechies 5, Meyer and Coiflet 5. Finally, we used a neural network MLP (multi layer perceptron) in order to improve the SVM results.

Digit recognition using wavelet and SVM in Brazilian Portuguese

This paper describes isolated word recognition of deaf students by unsupervised and supervised neural network. Compared to normal speech, there is high variability in deaf speech and by hearing once we couldn't understand it. By the use of proposed method deaf people can make use of all voice operated devices. In this paper we use combination of SOFM and BPN neural network for recognition. Initially the input is sampled, filtered, windowed and Perceptual Linear Predictive Coefficients are determined for each frame. These coefficients are applied as input to the SOFM neural network. The output of this network is given to BPN neural network comprising of 3 layers for learning. The network has been trained with five words uttered by five different deaf persons in the age group of 5–10 years. Another set of same five words uttered by same five deaf persons were used for test purposes. The recognition results for the word one, three, four is 50 to 60% and for five is 50%…But the recognition results for word two are only10% since the variability is high for two. The results can be improved by varying the parameters of the hybrid neural network.

Speech recognition of deaf and hard of hearing people using hybrid neural network

This thesis is concerned with the application of artificial neural network models to speaker verification. In particular, the application of unsupervised and supervised neural networks with text dependent isolated digit systems is addressed. A hybrid neural network is proposed for speaker verification (SV). It consists of self organized neural networks and a network of multi-layer perceptrons (MLP). The basic idea in this system is the usage of vector quantization preprocessing as the feature extractor. The preprocessing stage measures local spectral similarities and uses this information to train the MLP. The benefits of this SV system are its speed and simplicity. The system was used in a series of experiments. The first experiment used a neural network model (NNM) with frame labelling performed by a client codebook known as NNM-C. Enhanced performance was obtained from this model when compared with the Hidden Markov Model (HMM) approach. The second set of experiments used the NNM with frame labelling from client and impostor codebooks known as NNM-CI. The results were poorer than when compared with the NNM-C and HMM. The amount of data, the choice of neural network parameters such as learning rate and momentum term and the use of hidden nodes are discussed. In the third experiment a new approach used a speaker verification (HMM-MLP) method which combines a HMM based preprocessor with MLP. The output scores of the HMM were used as the inputs to the MLP. Compared with the NNM-C it is more successful since it has the capability of time alignment of the speech signals together with the efficient discrimination capability of the neural network. Moreover the method achieves better performance by the use of more than one feature set for each set of preprocessed parameters. It was found that further improvements in verification performance was achieved with the addition of the MLP and the HMM-MLP combination achieved 25% equal error rate reduction compared to the traditional HMM. Finally, the most important contribution of this research was the development and refinement of a NNM SV approach which incorporates a client barcode into the system design. The conclusion from the experiments is that use of optimised preprocessors, especially HMMMLP, yields improved results but this must be weighed against the increased complexity of system design.

Evaluation of preprocessors for neural network speaker verification

A hybrid neural network is described. It consists of a Kohonen map and a perceptron. The hybrid is proposed firstly for speaker independent, isolated word recognition. However, it may also be used for other classification problems. The novel idea in this system is the usage of a Kohonen map as the feature extractor which converts phonetic similarities of the speech frames into spatial adjacency in the map. This property simplifies the classification task. The system performance was evaluated for recognition of a limited number of Farsi words (numbers "zero" through "ten"). The overall performance of the recognizer showed to be 93.82%. >

S.B. Zahirazami

Papers

Isolated word recognition using a hybrid neural network